Single-clip overfit (S2 · cell D): anchor gap 0.071 m, no micro-stepping, eyeball-verified ~70–80% fidelity.
Left = reference data, right = overfit policy. Gap 0.071 m — robot and anchor ball move as one. Passed visual review.
2 · Why it matters
This is the ceiling proof for the whole project: SONIC's stack can reproduce reference motion on the X2 embodiment to near-reference fidelity. Every diagnosis since anchors on it:
Anything the overfit policy can do that the fleet policy can't is a training problem — never blamed on the robot.
The residual ~20–30% (slightly different flavor, no micro-steps) is consistent with the three embodiment boundaries (shoulder, leg length, motor speed — see the boundaries report).
SBS clips: left = reference motion data, right = policy rollout (Isaac Lab, full physics) · yellow ball = reference anchor · all runs single RTX 3090